Developing a prediction model of neonatal pneumonia in patients with gestational diabetes mellitus in primary medical institutions

Abstract

Background The incidence of gestational diabetes mellitus (GDM) continues to rise in China, and the risk of neonatal pneumonia is significantly increased, but the primary medical institutions lack simple and effective early risk assessment methods. Objective To construct a prediction model for neonatal pneumonia suitable for primary medical institutions, based on readily available clinical characteristics of pregnant women with GDM, we aimed to identify key risk factors and provide a basis for developing clinical intervention strategies. Methods Data on pregnant women and their newborns at Anyue Maternal and Child Health Hospital from 2023 to 2025 were retrospectively collected. Eight machine learning methods were used to construct the prediction model by bootstrap method and SHapley Additive exPlanations (SHAP) was used to enhance the model's interpretability. Results A total of 328 pairs of mothers and newborns were included, and the incidence of neonatal pneumonia was 8.8%. Among the eight prediction models, Random Forest model performed the best and Decision tree model was the worst; SHAP results indicated education level, BMI, hyperemesis gravidarum, primipara, hyperlipidemia during pregnancy, premature birth and job were most influencing factors. Binary logistic regression analysis showed that premature birth (OR = 4.763, 95% CI = 1.522−14.901, P = 0.007), Hyperemesis Gravidarum (OR = 3.706, 95% CI = 1.587−8.653, P = 0.002), and cesarean section (OR = 2.456, 95% CI = 1.050−5.745, P = 0.038) were risk factor of neonatal pneumonia, with a protective factor of higher education level (OR = 0.370, 95% CI = 0.158–0.865, P = 0.022). Conclusion Despite the limited number of neonatal pneumonia events, our findings provide preliminary evidence that it may be feasible to develop a simplified risk-stratification tool for neonatal pneumonia in pregnancies complicated by GDM using routinely available clinical variables. Further multicentre prospective studies are required to determine the optimal predictor set and externally validate the model before clinical application.